Two-level hierarchical bayesian poisson models for small area estimation of infant mortality rates

Abstract

Hierarchical Bayes (HB) approach has been recently proposed for small area estimation problems. However, in this area, there is still a limited use of hierarchical Bayesian models though they have many advantages: (i) their specification is straightforward and allows to take into account the different sources of variation and (ii) inferences are clear-cut and computationally feasible in most cases by using standard Markov Chain Monte Carlo (MCMC) techniques. Within this approach, when the variable of interest is a count or a proportion, alternative model specifications can be considered. In the HB approach, a prior distribution on the model parameters is specified and the posterior distributrion of the parameters of interest is obtained. Inferences are based on the posterior distribution, in particular, a parameter of interestis estimated by its posterior mean and its precision is measured by its posterior variance. Even though the HB is straightforward, and HB inferences are clear-cut and „exact‟, but the HB requires the specification of a subjective prior ( ) on the model parameter . Priors on maybe informative or „diffuse‟. Informative priors are based on substantial prior information, such previous studies judged relevant to the current data set y. However, informative priors are seldom available in real HB application, particularly those related to public policy. This dissertasion analyses and discusses the structure of alternative models when the variable of interest is a count, both on a theoretical side as well as by a simulation study. Bayesian specifications derived from classical models for SAE, e.g. the Fay-Harriot model (Fay and Herriot, 1979), and then consider a generalized linear Poisson model. The combination of these two models has formed a model called the hierarchical Bayes Poisson regression model. Development of HB models that extend the Fay-Herriot models and generalized linear model by using two different priors distribution on the model parameter of interest. First, we used gamma distribution as a conjugate priors for Poisson likelihood, and second we used inverse gamma distribution which is non conjugate prior for Poisson likelihood. The proposed models are implemented using the Gibbs sampling method for fully Bayesian inference. We apply the proposed models to the analysis of infant mortality rate for sub-district level in Bojonegoro district, East Java Province. Based on the application result, we found that hierarchical Bayes Poisson regression model with inverse-gamma prior distribution gave the better prediction of infant mortality rate than model with gamma prior distribution. Our conclusion were made based on various criterion on diagnostic models, such as convergence of Markov chain diagnostic tests, a Bayesian measure of fit or adequacy, residual analysis, and other summary statistics (bias and mean square error).